Merge tokens expands C by merge^2 (merger takes pre-expanded)

This commit is contained in:
Daniel Maddern 2026-08-19 22:36:32 +07:00
parent 571d3541e8
commit b085ef02e8

View file

@ -681,7 +681,7 @@ class Qwen3VL32BVision(nn.Module):
C = int(x.shape[-1])
merge = 2
target = (t, h // merge, merge, w // merge, merge, C)
return x.view(target).permute(0, 1, 3, 2, 4, 5).reshape(-1, C // (merge * merge))
return x.view(target).permute(0, 1, 3, 2, 4, 5).reshape(-1, C * merge * merge)
def forward(self, flatten_patches: torch.Tensor, grid_thw: torch.Tensor) -> tuple[torch.Tensor, list[torch.Tensor]]:
"""Run the visual tower -> (merged, deepstack)."""